Variational autoencoders for anatomy-specific insights in medical imaging: A systematic review

Document Type

Review

Publication Date

5-1-2026

Abstract

Variational Autoencoders (VAEs) have emerged as powerful generative models in medical imaging due to their ability to learn compact latent representations and model data distributions under limited supervision. Despite their growing applications, a systematic understanding of how VAEs are used to capture anatomy-specific structural characteristics, such as organ morphology, pathological variations and structural anomalies remains underexplored. This systematic review aims to analyze and summarize the role of VAEs in structural medical imaging, providing insights into their effectiveness for tasks such as segmentation, classification, anomaly detection, and image synthesis across various anatomical areas. The primary agenda is to provide a comprehensive taxonomy of how VAEs are engineered to address the unique morphological constraints of different organs. Following PRISMA guidelines, we conducted a systematic synthesis of 118 studies identified via the Web of Science. The reviewed works are categorized by anatomical domains, including neuroimaging, cardiac imaging, ophthalmic imaging, and lung imaging, with emphasis on imaging modalities and downstream tasks. VAEs are most extensively applied in neuroimaging, particularly for MRI-based synthesis and anomaly detection. Other domains exhibit distinct modelling priorities driven by temporal dynamics, intensity-based patterns, or fine-scale structural variability. Key challenges include limited generalizability, training specifications, heterogeneous evaluation practices, and privacy-related constraints. VAEs are powerful frameworks for anatomy-aware representation learning, but their clinical translation requires standardized benchmarking and multimodal integration. The review highlights the potential of VAEs to advance precision medicine by improving computer-aided tasks and enabling efficient model training from limited data.

Keywords

Deep learning, Generative AI, Healthcare, Medical imaging analysis, Variational autoencoders

Publication Title

Applied Soft Computing

ISSN

1568-4946

DOI

10.1016/j.asoc.2026.114913

Volume

194

Publisher

Elsevier

Link to Full Text

Share

COinS